Denetimsiz anomali tespit algoritmaları

Detection of outliers or anomalies in the data is of great importance in data analysis. Different approaches can be used in anomaly detection according to type of the problem. Unsupervised anomaly detection (UAD) approach is the most challengeable part of these approaches. UAD methods aim to detect anomalies without using a labelled training dataset. UAD algorithms can be considered in three main groups: nearest neighbour based, clustering based and statistical based. In this thesis, UAD approaches is examined and an adjustment, that depends on sample size, is proposed for statistical based algorithm, HBOS. In the first part of the application, performance of the most widely used UAD algorithms, that are k-nearest neighbour (k-NN), local outlier factor (LOF), local density cluster-based outlier factor (LDCOF) and histogram-based outlier score (HBOS), are compared. According to the results, HBOS algorithm is found more successful in terms of accuracy rate and runtime. In the second part of the application, effect of the bin-width determination techniques on to the performance of HBOS algorithm is examined. According to the results of the comparison with multivariate data of different characteristics, there is no superiority between the bin-width determination techniques in terms of accuracy.